Math Problems · AI-MO
NuminaMath-1.5
860K+ competition math problems from 17 sources with verified solutions — the training backbone of the gold-medal solution at the 2024 AI Mathematical Olympiad.
What NuminaMath-1.5 is
NuminaMath-1.5 is an English-language post-training mathematical dataset published by AI-MO. The dataset served as the training backbone for the gold-medal winning entry in the 2024 AI Mathematical Olympiad progress prize.
The collection addresses the challenge of providing verifiable, structured mathematical reasoning data for training language models. Materials span Chinese high school exercises, national contests, online forum discussions, and international Olympiad competitions, accompanied by detailed step-by-step solutions.
For computational researchers, NuminaMath-1.5 provides enriched problem-level metadata including categorised mathematical domains, problem formats, and isolated ground-truth answers to facilitate automated verification and evaluation.
How it was made
The dataset compiles mathematical problems gathered primarily from online examination paper PDFs, official national Olympiad websites, and mathematics discussion forums across 17 sources. Each item includes a natural language problem statement alongside a chain-of-thought reference solution explaining the steps required to resolve it.
NuminaMath-1.5 adds standardised metadata fields: 'problem_type' categorises topics across algebra, geometry, number theory, combinatorics, calculus, inequalities, logic and puzzles, or other; 'question_type' classifies items into multiple-choice questions, proofs, or math-word problems; and 'answer' extracts numerical values, denotes proofs, or flags missing outputs as 'notfound'. Data collection also incorporates manual parsing and curation to resolve earlier text-extraction errors from regular expressions and models, while unhelpful synthetic sets such as 'synthetic_amc' were pruned.
Key facts
- Total problems
- 896,215
- Format
- Parquet
- Language
- English
- Modality
- Text
- Release year
- 2024
- Problem types
- Algebra, Geometry, Number Theory, Combinatorics, Calculus, Inequalities, Logic and Puzzles, Other
- Question types
- Math-word-problem (631,522), multiple-choice question (146,449), proof (110,998)
What you can use it for
- Supervised fine-tuning for mathematical reasoning
- Train generative language models on multi-step chain-of-thought solution trajectories across diverse mathematical domains.
- Automated reasoning evaluation
- Test model accuracy against explicit numerical answers or proof structures annotated with problem type metadata.
- Domain-specific model ablation
- Filter and experiment on focused mathematical subfields, such as inequalities, geometry, or number theory, using subset tags.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from AI-MO/NuminaMath-1.5 at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 896,215 rows of default/train
answer
string_text- min
- 0
- median
- 5
- mean
- 9.11
- max
- 575
5.7% empty
problem
string_text- min
- 0
- median
- 216
- mean
- 262.35
- max
- 45,303
problem_is_valid
string_label- Yes96%
- Incomplete3%
- NaN1%
problem_type
string_label- Algebra47%
- Geometry21%
- Number Theory10%
question_type
string_label- math-word-problem70%
- MCQ16%
- proof12%
solution
string_text- min
- 0
- median
- 807
- mean
- 965.03
- max
- 26,844
3.6% empty
solution_is_valid
string_label- Yes94%
- Incomplete4%
- Problem not solved2%
source
string_label- cn_k1230%
- olympiads22%
- orca_math17%
synthetic
bool- False65%
- True35%
Strengths and limits
Good at
- Features comprehensive problem-level metadata including ground-truth answers, question types, and topic categories.
- Contains manually parsed and verified Olympiad problems sourced directly from official national competition websites to avoid automated parsing errors.
- Supplies 896,215 diverse competition problems ranging from secondary school exercises to elite international Olympiads.
Watch out for
- Contains solutions where the final numerical answer could not be extracted and is labelled as 'notfound'.
- Proof-style questions cannot be automatically checked with a standard scalar answer value.
- Certain synthetic data generation techniques were found to degrade performance, prompting the removal of datasets like 'synthetic_amc'.
Questions
Can I look inside NuminaMath-1.5 without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is NuminaMath-1.5 under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What mathematics disciplines are covered in NuminaMath-1.5?
The dataset categorises problems into algebra, geometry, number theory, combinatorics, calculus, inequalities, logic and puzzles, and a general other category.
How are problem answers represented?
For math word problems with numerical outputs, the final answer is provided directly. For proofs, the answer field is labelled 'proof', and if an answer cannot be parsed from the reference solution, it is marked 'notfound'.
How does NuminaMath-1.5 differ from the earlier NuminaMath-CoT dataset?
Version 1.5 adds structured metadata for problem types, question types, and final answers, incorporates manually curated and parsed Olympiad and contest sets, and removes synthetic data subsets that harmed model performance.
What formats are the mathematical problems presented in?
Problems are divided into three formats: math word problems (631,522 problems), multiple-choice questions (146,449 problems), and mathematical proofs (110,998 problems).
How to cite it
Jia LI, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, Stanislas Polu. NuminaMath. 2024. https://github.com/project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf
Written from NuminaMath-1.5's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.